即老数据中的image为单个频段，新数据中image为多个频段，可以自动判断并读取。
使用时需要传入两个不同的数据路径:train_dict_root以及train_dict_root1，即支持多个数据集一块训练，其中train_dict_root1可以为None
train_dict_root = "/data/xbj/0927_500hz_with_clean/train"
train_dict_root1 = "/data2/hsl/0324_pt_with_clean" # 新数据地址
代码中具体调用方式为：
train_transformed_dataset = Dataset_dict(root_dir=train_dict_root,root_dir1=train_dict_root1, transform=data_transform,
start_freq=args.start_freq, end_freq=args.end_freq, which_freq=args.which_freq, failed_file=failed_file)
此外，还需更改训练的频率，各个频率对应为 125Hz:(0, 7); 250Hz:(1, 10); 500Hz:(2, 13); 1kHz:(3, 16); 2kHz:(4, 19).
parser.add_argument('--which_freq', type=int, default=2) # 记得改成对应的图编号
parser.add_argument('--freq_loc', type=int, default=13) # 记得改成对应的倍频程
python train_resnet50_500Hz_multi_card.py
python -m torch.distributed.launch --nnodes=1 --nproc_per_node=3 train_resnet50_1000Hz_multi_card_DDP.py
其中nnodes为机器数，不用改，nproc_per_node为显卡数，更快更推荐，使用时记得更改数据集路径及选择频率，步骤与上面一样
train_resnet50_1000Hz_multi_card_DDP_fixed.py 修补了train_resnet50_1000Hz_multi_card_DDP.py运行的一些问题，在运行程序时，采用train_resnet50_1000Hz_multi_card_DDP_fixed.py
new_data_load_original_fixed.py 修补了new_data_load_original.py运行的一些问题，在运行程序时，采用new_data_load_original_fixed.py
可以批量测试，比较简单，详见代码
test_resnet50_500Hz.py
The *.py is used to test previous data, and it can test different frequency. You can run your own test by the Run Instructions.
If you want to test the newly generated data, you should revise the following place:
1、The default of the following two codes should be modified to the path of your pretrained model.
parser.add_argument('--save_dir', type=str, default="/data2/pyq/yousonic/old_500hz/1001_autoLoss/save_model/")
parser.add_argument('--model_path', type=str, default="/data2/pyq/yousonic/old_500hz/1001_autoLoss/save_model/")
2、The default represents the number of epochs.
parser.add_argument('--epoch_for_save', type=int, default=60)
3、The defaults are used to choose frequency, and the corresponding relationships are 125Hz:(0, 7); 250Hz:(1, 10); 500Hz:(2, 13); 1kHz:(3, 16); 2kHz:(4, 19).
parser.add_argument('--which_freq', type=int, default=2) # 记得改成对应的图编号
parser.add_argument('--freq_loc', type=int, default=13) # 记得改成对应的倍频程
4、The default should be modified to your own save-path.
parser.add_argument('--outputresult_dir', type=str, default="./0329_oldmodel_yqhdata_500hz/epoch60")
5、The val_dict_root and failed file should be modified to your own data path and you can revise the val_batch_size. However your data must have clean !!!
if DEBUG == False:
val_dict_root = "/Users/bajianxiang/Desktop/internship/filter_down_spec_dataset"
val_batch_size = 1
failed_file = "/Users/bajianxiang/Desktop/internship/filter_down_spec_dataset/koli-national-park-winter/koli-national-park-winter_koli_snow_site4_1way_bformat_1_koli-national-park-winter_东北话男声_1_TIMIT_a005_100_110_10dB-0.pt"
else:
val_dict_root = "/data/xbj/0927_500hz_with_clean/val"
failed_file = "/data/xbj/0927_500hz_with_clean/val/hoffmann-lime-kiln-langcliffeuk/hoffmann-lime-kiln-langcliffeuk_ir_p2_0_1_hoffmann-lime-kiln-langcliffeuk_东北话男声_1_TIMIT_a053_110_120_10dB-0.pt"
val_batch_size = 6
6、If you use model train through multi-card, you should add the following code.
net = nn.DataParallel(net)
model
utils
data_load.py
data_load_aug2.py
data_load_for_FPN_spec.py
eval_same_room.py
gen_boxplot.py
get_deverb_spec_test.py
gtg.py
load_same_room.py
new_data_load_original.py
new_data_load_original_for_our_data.py
new_data_load_origin_old.py
new_thread_test.py
splweighting.py
SSIMLoss.py
Test_new_dataset.py
test_our_data.py
thread_test.py
train_resnet50_500Hz.py
valdata_meant60.py
workspace_utils.py